|
Abstract:
|
Large-scale spatiotemporal forecasting techniques have been widely adopted in various urban computing applications, such as transportation management, ride-sharing market and urban planning. There are several potential difficulties specifically appears in this research field to be tackled. First, large-scale multi-modality urban data should be represented and fused in efficient ways to improve the performance and availability of model. Second, the algorithm is required to handle unsmooth and non-stationary signals in both spatial and temporal dimensions. Third, trained model and knowledge are hard to be transferred among cities due to different city layouts, which increases the efforts of recalibrating model in new cities. Fourth, most existing forecasting methods select important variables via their correlations but not causalities with responses of interest, leading to poor interpretability. In this {paper/presentation}, we will introduce four research works to address the above issues. First, we represent spatial layouts of cities in graph formats and employ graph neural networks (GNN) to solve spatiotemporal forecasting problem. Second, we propose the clustered attention to ca
|
